Interview with Matt Ranney

QCon: What is the motivation for your talk?

Matt: The biggest problem Uber has is handling growth (product, business, and engineering). The team has gone from (when I first started) 200 engineers to 1,100 engineers. Handling that growth has been a huge problem. I don’t just mean people (that’s a huge problem too), I mean from technology viewpoint. How do you have all those engineers making meaningful contributions to your systems?

As a result, we left broken a bunch of stuff that would have been fixed if we were proceeding methodically. The motivation for my talk is about how we’re coping with adding this many people (this quickly) without taking the time to do a consistent design, it’s more adapting to chaos as much as taming it.

QCon: How is this talk different than the one you did in London earlier this year on Uber’s architecture?

Matt: The London talk was about the dispatch system. We went into detail into some of it. The dispatch system now runs the entire world. When we talked about it earlier this year, it was only in one city, but now it’s what ‘makes Uber go’.

In San Francisco, we’re talking about what we had to do on the infrastructure side to scale services. Things around routing and the explosion of of our services. Uber has about 500 services in production now. It’s a shocking graph anytime anyone sees it.

This talk is about the work we’re doing to get a handle on all that.

QCon: Can you give me an idea of some of the things you might be talking about?

Matt: There are some pretty neat failure handling techniques that I would like to go into that we have in our system. Those are among the coolest.

For example, we implement in our RPC framework Backup Requests w/ Cross-Server Cancellation from Jeff Dean’s presentation a year or two ago. As far as I know, we are the only system that does that. It’s an interesting technique. We use the same technique (basically making everything you can be Idempotent) everywhere we can, so that a service can be retried (or run twice accidentally) to mask failures as well as handle latency outliers.

QCon: Who do you think are the people that should be most interested in your talk?

Matt: If people are interested in what are the problems of scaling a rapidly growing engineering organization, the talk will be interesting to you. There are no specific technologies you need to know to come to the talk.